Distributed electric vehicle,Nonlinear model predictive contro,einforcement learning

Under extreme conditions such as low-adhesion surfaces or high-speed sharp turns, the sudden change in tire dynamics from the linear region to the saturation region can cause dynamic mismatch issues in traditional fixed-parameter controllers. This paper proposes an adaptive integrated control architecture combining Soft Actor-Critic (SAC) reinforcement learning with Nonlinear Model Predictive Control (NMPC).

The upper layer is based on a seven-degree-of-freedom vehicle model and a composite tire slip model to construct the NMPC framework, which integrates longitudinal and lateral control to calculate the additional yaw moment. The lower layer proposes a yaw moment dynamic optimization distribution algorithm for four wheels, considering slip rate tracking error, tire load rate, and torque distribution error. To address dynamic mismatch caused by severe changes in operating conditions, a Multi-Head Attention (MHA)-SAC algorithm is introduced to adaptively adjust the prediction horizon and cost function weights of NMPC, as well as optimize the weighting coefficients of the sub-objective functions in the lower layer.